Newer RDKit versions have changed so drastically that DiffDock's accuracy drops by ~50%. This happens because it suddenly sees features it was never trained on.
This commonly happens when using the Posebusters package in the same venv as DiffDock. The issue seems to happen to other people too and is explained in detail here.
A comment in the README or an assertion of the correct feature range, as in training, would save many hours of debugging for future users.
Newer RDKit versions have changed so drastically that DiffDock's accuracy drops by ~50%. This happens because it suddenly sees features it was never trained on.
This commonly happens when using the Posebusters package in the same venv as DiffDock. The issue seems to happen to other people too and is explained in detail here.
A comment in the README or an assertion of the correct feature range, as in training, would save many hours of debugging for future users.